Optimizing key value cache for large language model inference | Patent Application Number 19191984 | Patexia